Google DeepMind has published a thorough technical report describing how artificial general intelligence (AGI) could potentially threaten humanity — and what can be done to prevent it, writes ArsTechnica.
The 108-page report, led by DeepMind co-founder Shane Legg, categorizes key AGI risks into four categories: misuse, inconsistency, errors, and structural threats. While the public is still focused on current generative AI tools like ChatGPT, DeepMind researchers argue that we could see human-level AI as early as 2030, and that the world is not ready for it.
The biggest concern is misuse—the potential for attackers to use AGI for malicious purposes. Unlike current AI tools, a powerful AGI system could detect software vulnerabilities, develop biological weapons, or organize large-scale cyberattacks. The authors warn that without robust protections, even a single person with access to such systems could cause enormous damage.
DeepMind offers extensive testing before deployment, enhanced post-training security, and even a method known as “unlearning” to remove dangerous capabilities from AGI. But the usefulness of these safeguards remains unproven.
More theoretical, but potentially more catastrophic, is the risk of imbalance, where AGI acts contrary to human values or instructions, despite appearing to be cooperative. This is the classic scenario of “rogue AI,” where an AI circumvents its limitations and acts in pursuit of its own goals.
To prevent this, DeepMind recommends methods such as enhanced supervision, where multiple AI systems evaluate each other's behavior, and safe sandboxes under human supervision. The idea is to detect early signs of autonomous behavior before the systems are widely deployed.
There are also unintentional errors that could increase as AGI capabilities scale. Modern AI systems often make factual errors, but DeepMind warns that mistakes made by AGI — especially in a military context or in the field of critical infrastructure — could have serious consequences.
They recommend taking it slow, keeping AGI on a short leash, and using filtering mechanisms ("shields") to intercept dangerous actions before they are executed.
Perhaps the most worrying are structural risks: unintended societal transformations brought about by the widespread deployment of AGI. These include the spread of hyper-realistic disinformation, economic destabilization, or the slow erosion of the human factor as machines quietly take over decision-making in politics and business.
These risks, DeepMind acknowledges, are the hardest to mitigate—and they may arise not from attackers or faulty code, but from systems functioning as intended. In other words, even well-intentioned AGI could change society in ways we can’t predict.
While opinions in the AI community are divided, DeepMind sees AGI as a possibility in the near future. Some researchers believe AGI is still decades away—or even impossible to create with current AI development approaches —but others, including leaders at OpenAI and Google, believe it could be here within this decade.
Tulse Doshi, a product lead on Google’s Gemini team, recently said that definitions of AGI vary widely. But the trajectory of smarter large language models is clearly moving toward higher intelligence and more complex reasoning—whether this amounts to “general intelligence” remains to be seen.
DeepMind’s authors are careful to see their work as a starting point, not a solution. But in a climate of rising geopolitical tensions, increasing investment in AI, and public excitement, they emphasize that it is time to have serious discussions about how to govern AI, contain it, and, if necessary, shut it down.
After all, if they are right, and AGI does indeed appear by 2030, the world will only have a few years left to create the necessary safety safeguards.